Losing Localization. Please help

[vimalrajayyappan]

Dear friend,

Thanks for the package.
I had tried using your package for localization where I found some drifts in localization even when Im fusing with IMU.

I used DIRECT7 method for registrationa and localization
NDT_OMP

I saved map at 0.01 resolution

when I try to localize i found some warning like
"Leaf size is too small for the input dataset. Integer indices would overflow."

I altered downsample_resolution and checked as well. Couldnt figure.I attached a localization config file screenshot.Please check.

Kindly help me to gain the accuracy.
Suggest me some steps.

Thanks a lot

[koide3]

Hi @vimalrajayyappan ,
Maybe your map is too large for voxelgrid filtering. Change VexelGrid in the following line to ApproximateVoxelGrid which can handle smaller resolutions.

hdl_localization/apps/globalmap_server_nodelet.cpp

Line 58 in 624a761

boost::shared_ptr<pcl::VoxelGrid<PointT>> voxelgrid(new pcl::VoxelGrid<PointT>());

[vimalrajayyappan]

Thanks a lot Koi. I will give a try.
Following which can you please share some values of params in Mapping&Localization and their insights. So that I can tune and check their optimality.

Another qn pls:
Should initial pose be highly accurate? Once I give the initial pose whether the algorithm tries to match the best position nearby, before I start moving? or it will just assumes the initial pose is its start position

[koide3]

The following guide would be helpful to tune the parameters.
https://github.com/koide3/hdl_graph_slam#the-mapping-result-is-poor-parameter-tuning-guide

The initial pose should be close to the actual pose (in particular, a large rotational error may affect the convergence). Once you give the initial pose, it updates the estimated pose even if the sensor is fixed.

[vimalrajayyappan]

Thanks a lot Koi... 😎Will keep updated

[fatmanurgezmiss]

ctual pose (in particular, a large rotational error may affect the convergence). Once you give the initial pose, it updates the estimated pose even if the sensor is fixed.

The following guide would be helpful to tune the parameters. https://github.com/koide3/hdl_graph_slam#the-mapping-result-is-poor-parameter-tuning-guide

The initial pose should be close to the actual pose (in particular, a large rotational error may affect the convergence). Once you give the initial pose, it updates the estimated pose even if the sensor is fixed.

Hi,
I have the same issue that called "Leaf size is too small for the input dataset. Integer indices would overflow." Then I saw yor answer for that. When I try ApproximateVoxelGrid terminal issue gones but localization fail is continues. Can you help me for that?